Beruflich Dokumente
Kultur Dokumente
Communication Monographs
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/rcmm20
To cite this article: Andrew F. Hayes (2009): Beyond Baron and Kenny: Statistical Mediation Analysis
in the New Millennium, Communication Monographs, 76:4, 408-420
To link to this article: http://dx.doi.org/10.1080/03637750903310360
Communication Monographs
Vol. 76, No. 4, December 2009, pp. 408420
Figure 1 The total effect of X on Y (A), a simple mediation model (B), a single-step
multiple mediator model (C), and a multiple-step multiple mediator model (D).
410 A. F. Hayes
In more complex models, as in Figure 1C or D, the same rules apply. In Figure 1C,
the total effect is equal to the direct effect of X on Y plus the sum of the indirect effect
through M and the indirect effect through W. That is, c c?a1b1a2b2. In a model
with two or more intervening variables, the indirect effect through a given
intervening variable is called a specific indirect effect (e.g., the specific indirect effect
of X on Y through M), and the sum of the specific indirect effects is called the total
indirect effect of X. In Figure 1D, the total effect of X on Y can similarly be partitioned
into direct and indirect components. Here, c c?a1b1a2b2a1a3b2, with the
latter three terms being specific indirect effects and their sum being the total indirect
effect (see Brown, 1997).
The Causal Steps Approach to Testing Intervening Variable Effects
Models such as these are frequently estimated by communication researchers.
Although there are many methods available for testing hypotheses about intervening
variable effects, the most widely-used method is the causal steps approach popularized
by Baron and Kenny (1986). This approach requires the researcher to estimate each of
the paths in the model and then ascertain whether a variable functions as a mediator
by seeing if certain statistical criteria are met. For example, if both a and b paths in a
model such as Figure 1B are statistically significant and c? is closer to zero than c, then
M is deemed a mediator of the relationship between X and Y. Some assess whether
ones data meet these criteria only if there is evidence of a total effect of X (i.e., if c is
statistically significant), one of the requirements of mediation outlined by Baron and
Kenny. If the significance of c is not used as a prerequisite to further examination of
the paths, then this causal steps approach is sometimes called a test of joint
significance.
Unbeknownst to many, the causal steps approach has been criticized heavily on
multiple grounds. Most notably, simulation studies have shown that among the
methods for testing intervening variable effects, the causal steps approach is among
the lowest in power (Fritz & MacKinnon, 2007; MacKinnon, Lockwood, Hoffman,
West, & Sheets, 2002). That is, if Xs effect on Y is carried in part indirectly through
intervening variable M, the causal steps approach is least likely of the many methods
available to actually detect that effect.Another criticism of this approach is that it is
not based on a quantification of the very thing it is attempting to test*the
intervening effect. Rather, the existence of an indirect effect is inferred logically by the
outcome of a set of hypothesis tests. If a and b are both different from zero by a
statistical significance criterion, then so too must be the indirect effect according to
the logic of this approach. But it is traditional in social science to base claims on tests
of quantities pertinent to those claims. We infer the existence of partial association,
differences between experimental conditions, and so forth by quantifying these effects
and then testing hypotheses about or constructing interval estimates for their size.
Given that an indirect effect is quantified as the product of its constituent paths,
should we not base inferences about indirect effects on tests of the product?
Additionally, it is possible for an indirect effect to be detectably different from zero
even though one of its constituent paths is not. Hypothesis tests are fallible. Each
carries with it a possibility of a decision error. The more nulls that must be rejected in
order to claim an indirect effect, the more likely the analyst will go away empty
handed. It makes more sense to minimize the number of tests one must conduct to
support a claim.
Given these problems with the causal steps approach, why is it so widely used? The
most plausible explanation is that it is simple and widely understood. Most anyone
can be taught this approach, its implementation described in only a few manuscript
lines, and readers and reviewers will be able to follow along without difficulty. Yet
these are not convincing reasons for the use of a method that is not optimal when
there are better alternatives.
Modernizing Our Thinking About and Testing of Indirect Influences
Modern Approaches to Inference about Intervening Variable Effects
New analytical opportunities arise if we quantify indirect effects rather than infer
their existence from a set of tests on their constituent paths. One inferential technique
is the product of coefficients approach, most well known as the Sobel test (Sobel,
1982, 1986). This test requires an estimate of the standard error of ab (see Preacher &
Hayes, 2004, for three estimators of the standard error). The ratio of ab to its
standard error is used as a test statistic for testing the null hypothesis that the true
indirect effect is zero, with the p-value derived from the standard normal
distribution.
Although the Sobel test enjoys some use, frequently it is used as a supplement to
the Baron and Kenny approach rather than instead of it. An investigator may first ask
whether the analysis meets the Baron and Kenny criteria for establishing mediation
and, if so, the analyst then conducts the Sobel test to attest to the validity of the
conclusions reached without it. There is little point to this exercise. The outcome of a
set of hypothesis tests about a and b is irrelevant and provides no additional
information beyond the Sobel test about the size or significance of the indirect effect.
Thus, one should not precondition the use of the Sobel test on significant paths
linking X to M or M to Y.
The Sobel test has a major flaw. It requires the assumption that the sampling
distribution of the indirect effect is normal. But the sampling distribution of ab tends
to be asymmetric, with nonzero skewness and kurtosis (Bollen & Stine, 1990; Stone &
Sobel, 1990). We should not be using tests that assume normality of the sampling
distribution when competing tests are available that do not make this assumption and
that are known to be more powerful than the Sobel test. Of the alternatives, two seem
to be winning the battle: bootstrapping, and the empirical M-test. Simulation
research shows that these methods tend to have highest power and the best Type I
error control. Although the empirical M-test (also known as the distribution of
products approach) is advocated by Holbert and Stephenson (2003) as the best
option available to media effects scholars (p. 566), it suffers the major weakness that
412 A. F. Hayes
estimate of the indirect effect itself, but unlike the Sobel test, it makes no assumptions
about the shape of the sampling distribution of the indirect effect, thereby getting
around this problem that plagues the Sobel test. Additionally, notice that no standard
error is needed to make the inference, rendering the controversy about how to best
estimate the standard error of the indirect effect moot. Finally, it is a very general
approach, in that it can be used for making inferences about indirect effects in any
intervening variable model, regardless of how complex and how numerous the paths
between X and Y.
Bootstrapping is being used with increasing frequency, although like the Sobel test,
it is sometimes reported as a supplement to the causal steps approach rather than
instead of it. I see no reason to report the results of both methods, although little
harm is done if inferences are based on the bootstrap results.
Can Effects that Dont Exist be Mediated?
If a mediator is a variable, M, that is causally between X and Y and that accounts at
least in part for the association between X and Y, then by definition X and Y must be
associated in order for M to be a mediator of that effect. According to this logic, if
there is no evidence that X affects Y, then how can Xs effect on Y be mediated and so
what is the point of estimating indirect and direct effects? But it is possible for M to
be causally between X and Y even if X and Y arent associated. In this case, some prefer
to avoid the term mediator when describing M and instead refer simply to Xs indirect
effect on Y through M (see Mathieu & Taylor, 2006, for a discussion of the distinction
between indirect effects and mediation).
The distinction between mediation and indirect effect is not always made by users
of the Baron and Kenny method, who may prematurely end the hunt for evidence of
indirect effects if there is no evidence that X and Y are associated. If the size of c
constrained the size of a and b and therefore their product, this logic would make
sense. Unfortunately, no such constraints exist, and it is easy to show that the claim
that X cant affect Y indirectly in the absence of a detectable total effect is false.
Consider, for example, the covariance matrix in Table 1 based on a sample of 100.
Suppose X is an experimental manipulation of exposure to political campaign news,
M a measure of trust in government, and Y a measure of the likelihood of voting in
the next election. A simple regression of Y on X yields a nonsignificant total effect of
X, c 0.147, p .20. It appears from this analysis that there is no relationship
Table 1 Variances (on the Diagonal) and Correlations (on the Off-diagonal) between
Four Variables in a Sample of Size 100
Y
X
M
W
1.381
0.063
0.171*
0.429***
0.252
0.388***
0.185
1.053
0.106
1.167
414 A. F. Hayes
between exposure to campaign news and likelihood of voting. And yet in a simple
mediation model, the indirect effect of X on Y through M is not zero by a 95% biascorrected bootstrap confidence interval based on 5000 bootstrap samples (0.463 to
0.016, with a point estimate of 0.209), as are the paths from X to M (a
0.794, pB.001) and M to Y controlling for X (b 0.264, p .035).3 These results
are consistent with the claim that exposure to campaign news reduces trust, which in
turn lowers the likelihood of voting. In this example, an investigator who proposed X
exerts its effect on Y indirectly through M would never have reached the stage of
testing the indirect effect if he or she insisted the total effect be detectably different
from zero.
That X can exert an indirect effect on Y through M in the absence of an association
between X and Y becomes explicable once you consider that a total effect is the sum
of many different paths of influence, direct and indirect, not all of which may be a
part of the formal model. For example, it could be that two or more indirect paths
carry the effect from X through Y, and those paths operate in opposite directions (cf.,
MacKinnon, Krull, & Lockwood, 2000). Much as a main effect in 2 2 ANOVA
might be nonsignificant if the simple effects are opposite in sign (i.e., a crossover
interaction), two or more indirect effects with opposite signs can cancel each other
out, producing a total effect and perhaps even a total indirect effect that is not
detectably different from zero, in spite of the existence of specific indirect effects that
are not zero.
Consider now a multiple mediator model of the same data in Table 1, in which two
variables M and W are both proposed as mediators of the effect of X on Y (as in
Figure 1C). Suppose W is a measure of perceived importance of the outcome of the
election. As noted earlier, the total effect of the manipulation (X) on likelihood of
voting (Y) is nonsignificant. However, 95% bias corrected bootstrap confidence
intervals (based on 5000 bootstrap samples) for specific indirect effects through trust
in government (X0M0Y) and perceived importance of the outcome of the election
(X 0W0Y) both do not include zero. The specific indirect effect of X on Y through
M is negative (0.473 to 0.063, with a point estimate of 0.226), whereas the
specific indirect effect through W is positive (0.006 to 0.436, with a point estimate of
0.191). The direct effect, as in the prior analysis, is not significant and in fact is almost
the same as the total effect (c?0.182, p.20), and all a and b paths are statistically
significant (albeit the X 0W path only marginally so, p .056).
In this example, exposure to campaign news appears to exert an effect on
likelihood of voting through two mechanisms that work in opposite directions, by
increasing perceived importance of the outcome, which then increases likelihood,
while simultaneously lowering trust in government, which reduces likelihood. These
two effects, operating together, in effect cancel each other out in the estimation of the
total effect and so appear as a nonsignificant total effect of the manipulation. The
moral of this example is that I concur with others (e.g., MacKinnon et al., 2000;
Shrout & Bolger, 2002) who recommend that researchers not require a significant
total effect before proceeding with tests of indirect effects. If you find a significant
indirect effect in the absence of a detectable total effect, call it what you want*
mediation or otherwise. The terminology does not affect the empirical outcomes. A
failure to test for indirect effects in the absence of a total effect can lead to you miss
some potentially interesting, important, or useful mechanisms by which X exerts
some kind of effect on Y.
Comparing Specific Indirect Effects
Researchers frequently test more complex models that include multiple linkages
between independent variables, proposed mediators, and outcomes. Some examples
in communication include Afifi, Afifi, Morse, and Hamrick (2008), Ledbetter (2009),
Smith, Downs, and Witte (2007), and Southwell and Torres (2006). The inclusion of
multiple pathways to an outcome means that different theories can be pitted against
each other in a single model (i.e., Theory A might propose variable M1 functions as a
mediator of Xs effect on Y, whereas Theory B proposes M2 as the mediator) while
eliminating the problem of estimation bias that occurs when multiple mediators that
are intercorrelated are tested individually in simple mediation models using the
Baron and Kenny approach (see Preacher & Hayes, 2008a). Although communication
researchers acknowledge that variables typically produce effects through multiple
pathways, the reliance on the piecemeal causal steps approach means investigators
rarely take the extra step of statistically examining and testing for differences in the
relative sizes of specific indirect effects.
The quantification of indirect effects allows the investigator to answer such
questions as whether the specific indirect effect of X on Y through proposed Mediator
1 differs in size from specific indirect effect through proposed Mediator 2. Because
specific indirect effects from X to Y are free of the scale of measurement of the
intervening variables (see Preacher & Hayes, 2008a), they are comparable without
standardization or any other form of transformation. MacKinnon (2000) presents a
method for conducting such contrasts by deriving the standard error for differences
between specific indirect effects. Preacher and Hayes (2008a) discuss pairwise
contrasts in single-step multiple mediator models with any number of mediators
and provide SPSS and SAS routines for bootstrap-based inference. In SEM, contrasts
can be conducted by imposing equality constraints on products of paths and then
examining whether those constraints produce a better or worse fitting model (see
Preacher & Hayes, 2008a, for LISREL and Mplus code).
Combining Mediation and Moderation
Mediation refers to a sequence of causal relations by which X exerts its effect on Y by
influencing intervening variables. Moderation, a term sometimes confused with
mediation, describes a situation in which Xs effect on Y varies as a function of some
third variable M, the moderator variable. A moderated effect is typically modeled
statistically as an interaction between X and the moderator variable, frequently
quantified as the product of X and M. Moderation can help us to understand how a
process operates if the moderator places constraints on how or when that process can
function.
416 A. F. Hayes
418 A. F. Hayes
[2]
[3]
[4]
Although the partitioning of a total effect into direct and indirect components in the manner
described here does not require the assumption that the errors in estimation are
uncorrelated, such intercorrelation can bias parameter estimates and standard errors.
Correlated errors can result from, among other things, the exclusion of variables from the
model that are correlated with two or more included variables.
Otherwise, I agree wholeheartedly with the position Holbert and Stephenson (2003) take in
their statement that communication researchers should place much more emphasis on the
estimation and testing of indirect effects than they have in the past.
Bootstrapping requires the raw data rather than just a covariance matrix. Readers interested
in the raw data used in this example can contact me at hayes.338@osu.edu and I will gladly
send it.
Contrary to conventional wisdom, it is possible (although rare in practice) for a standardized
coefficient to be larger than 1 in absolute value (Deegan, 1978). This means that even the
standardized indirect effect has no real upper or lower bound.
References
Afifi, T. D., Afifi, W. A., Morse, C. R., & Hamrick, K. (2008). Adolescents avoidance tendencies and
physiological reactions to discussions about their parents relationship: Implications for
postdivorce and nondivorced families. Communication Monographs, 75, 290317.
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social
psychological research: Conceptual, strategic, and statistical considerations. Journal of
Personality and Social Psychology, 51, 11731182.
Bauer, D. J., Preacher, K. J., & Gil, K. M. (2006). Conceptualizing and testing random indirect effects
and moderated mediation in multilevel models: New procedures and recommendations.
Psychological Methods, 11, 142163.
Bollen, K. A., & Stine, R. (1990). Direct and indirect effects: Classical and bootstrap estimates of
variability. Sociological Methodology, 20, 115140.
Brown, R. L. (1997). Assessing specific mediational effects in complex theoretical models. Structural
Equation Modeling, 4, 142156.
Cheong, J., MacKinnon, D. P., & Khoo, S. T. (2003). Investigation of mediational processes using
parallel process latent growth modeling. Structural Equation Modeling, 10, 238262.
Cole, D. A., & Maxwell, S. E. (2003). Testing mediation models with longitudinal data: Questions
and tips in the use of structural equation modeling. Journal of Abnormal Psychology, 112,
558577.
Deegan, J. R. (1978). On the occurrence of standardized regression coefficients greater than 1.
Educational and Psychological Measurement, 38, 873888.
420 A. F. Hayes
Preacher, K. J., & Hayes, A. F. (2004). SPSS and SAS procedures for estimating indirect effects in
simple mediation models. Behavior Research Methods, Instrumentsand Computers, 36,
717731.
Preacher, K. J., & Hayes, A. F. (2008a). Asymptotic and resampling strategies for assessing and
comparing indirect effects in multiple mediator models. Behavior Research Methods, 40,
879891.
Preacher, K. J., & Hayes, A. F. (2008b). Contemporary approaches to assessing mediation in
communication research. In A. F. Hayes, M. D. Slater, & L. B. Snyder (Eds.), The Sage
sourcebook of advanced data analysis methods for communication research (pp. 1354).
Thousand Oaks, CA: Sage.
Preacher, K. J., Rucker, D. D., & Hayes, A. F. (2007). Assessing moderated mediation hypotheses:
Theory, method, and prescriptions. Multivariate Behavioral Research, 42, 185227.
Scheufele, D. A. (2002). Examining differential gains from mass media and their implications for
participatory behavior. Communication Research, 29, 4665.
Shrout, P. E., & Bolger, N. (2002). Mediation in experimental and nonexperimental studies: New
procedures and recommendations. Psychological Methods, 7, 422445.
Smith, R. A., Downs, E., & Witte, K. (2007). Drama theory and entertainment education: Exploring
the effects of a radio drama on behavioral intentions to limit HIV transmission in Ethiopia.
Communication Monographs, 74, 133153.
Sobel, M. E. (1982). Aysmptotic confidence intervals for indirect effects in structural equation
models. In S. Leinhardt (Ed.), Sociological methodology (pp. 290212). San Francisco:
Jossey-Bass.
Sobel, M. E. (1986). Some new results on indirect effects and their standard errors in covariance
structure models. In N. Tuma (Ed.), Sociological methodology (pp. 159186). Washington,
DC: American Sociological Association.
Southwell, B. G., & Torres, A. (2006). Connecting interpersonal and mass communication: Science
news exposure, perceived ability to understand science, and conversation. Communication
Monographs, 73, 334350.
Stone, C. A., & Sobel, M. E. (1990). The robustness of total indirect effects in covariance structure
models estimated with maximum likelihood. Psychometrika, 55, 337352.
Williams, J., & MacKinnon, D. P. (2008). Resampling and distribution of the product methods for
testing indirect effects in complex models. Structural Equation Modeling, 15, 2351.